core/ops/fft: Fft pulsification on non-FFT axes + guards

[JulienBalianSonos]

Summary

Two commits on this branch.

1. Fft::axes_mapping = natural. Fft is rank-preserving and 1-to-1 on every axis (only the values along the FFT axis change), so the generic per-pulse wrapper can handle streaming on any non-FFT axis once axes_mapping reports the 1-to-1 relationship. Without this, pulse-pass bails on No specific pulse transformation for Fft, and could not track pulsing axis, which blocks STFT-based streaming because STFT pulse-lowers to a streaming-axis frame loop with a per-frame Fft running on a different axis (the frequency axis).

2. fft_pulsify: reject streaming on the FFT axis itself or on the trailing (re, imag) axis. The natural mapping would otherwise let PulseWrappingOp happily run Fft per-pulse on the FFT axis. That's a coherent operation: it's equivalent to Stft { frame: pulse_size, stride: pulse_size, axis } where the runtime picks pulse_size. But that is not the typed-model semantics the user wrote (Fft { axis } means "FFT every sample on that axis", which requires the axis length to be known and is undefined when the axis is symbolic). Silently reinterpreting it would be a foot-gun, so the pulsifier bails with a message that points users at Stft { frame: N, stride: N } and asks them to pick N explicitly. The complex-axis case is already rejected upstream in Fft::output_facts (trailing dim must be 2); the test locks in that earlier-layer rejection.

Edge-case sweep in harness/core-proptest-pulse/src/fft.rs:

• Smoke + proptest for forward FFT on rank-4 (batch, stream, fft, 2).
• Smoke + proptest mirror for inverse FFT.
• Rank-3 (no batch) and rank-5 (extra leading dims, matches DPDFNet df_op shape) smoke cases.
• Stacked FFT→iFFT roundtrip on the same axis (streaming traverses both via the natural mapping).
• Negative paths: streaming on the FFT axis bails cleanly; streaming on the complex axis is rejected at typed-model build.

Status

Originally opened with just commit 1, self-closed pending the edge-case work the comment thread called out. Re-opened now with commit 2 covering the guards and the wider test sweep.

[kali]

Did you try just implementing axes_mapping ? Not that the propose implementation is incorrect. It can't be ::natural() the time axis and (and the complex axis) do not map.

I think with the map right, there should not be any need for dedicated pulsification.

[JulienBalianSonos]

@kali I tried to fix according to your remark to make it more principled. let me know